A calibration coverage table answers a blunt question: when a tool draws a “90% confidence” band, do 90% of real outcomes actually land inside it? Reading one takes five minutes and protects you from most overconfident products.
A coverage table has one row per forecast horizon (say 3 months, 6 months, 1 year) and compares two columns: the stated coverage the tool claims (e.g. 90%) and the actual coverage it achieved — the share of past forecasts whose real outcome landed inside the stated band. If a 90% confidence interval is honest, those two numbers should be close. The gap between them is the whole story.
Actual well below stated (a “90%” band covering only 75% of outcomes) means the tool is overconfident: its bands are too narrow and reality escapes them more often than promised. This is the dangerous direction — you’ll be blindsided more than the number led you to expect.
Actual well above stated (a “90%” band covering 98%) means the tool is under-confident: its bands are wider than they need to be. Safer, but the forecast is vaguer than advertised — a band wide enough to always be right tells you very little. Actual close to stated (within a few points of 90%) is the goal: the tool’s confidence means what it says.
Here’s what this looks like with real, committed numbers. On Quantustik’s committed TOP-20 backtest, measured against the model as served, the 90% confidence bands have actually covered roughly 90.6% of outcomes at the 3-month horizon, 90.9% at 6 months, and 90.2% at 1 year. All three land within a point of stated — the “close to stated” case described above, which is what a healthy aggregate row looks like. Crucially, these are aggregate numbers, and the aggregate is the flattering part: coverage varies widely by individual stock, and both failure directions from the table above show up at the stock level. AVGO’s 1-year band covered only about 36.5% of outcomes — badly over-confident — while several names sit near 99%, which is the under-confident, too-vague end. A tidy aggregate row can sit on top of both. That stock-by-stock honesty — published on our live calibration page — is exactly what you should demand from any tool before trusting its confidence.
Confidence bands are how a forecast tells you how much to trust it. If the bands are overconfident, you’ll size positions as if the future is more certain than it is — the fastest way to get hurt by a “high-confidence” call that fails. A coverage table is the receipt that proves the confidence is real. A tool that shows you one, misses included, has earned a look; one that only quotes a single headline accuracy number has not.
This lesson is investor education, not personalized advice. A calibration table describes past coverage across many forecasts; it is not a promise about any single future outcome or a promise of profit.
For each forecast horizon it shows the tool’s stated confidence (e.g. 90%) next to the actual share of past outcomes that landed inside that band. If they’re close, the confidence is honest; a big gap reveals over- or under-confidence.
75% is the more dangerous case — the band is overconfident and too narrow, so reality escapes it far more often than promised, and you’d size positions as if the future were more certain than it is. 98% is under-confident: safer, but the forecast is vaguer than advertised.
Quantustik publishes one on its live calibration page, including how coverage varies stock by stock and where the model is weak. That kind of misses-included disclosure is the standard to demand from any tool before trusting its confidence.
Investor education only — not investment advice, and never a promise of profit. Every investment can lose value.